A regret lower bound for assortment optimization under the capacitated MNL model with arbitrary revenue parameters

Abstract In this note, we consider dynamic assortment optimization with incomplete information under the capacitated multinomial logit choice model. Recently, it has been shown that the regret (the cumulative expected revenue loss caused by offering suboptimal assortments) that any decision policy endures is bounded from below by a constant times $\sqrt {NT}$, where $N$ denotes the number of products and $T$ denotes the time horizon. This result is shown under the assumption that the product revenues are constant, and thus leaves the question open whether a lower regret rate can be achieved for nonconstant revenue parameters. In this note, we show that this is not the case: we show that, for any vector of product revenues there is a positive constant such that the regret of any policy is bounded from below by this constant times $\sqrt {N T}$. Our result implies that policies that achieve ${{\mathcal {O}}}(\sqrt {NT})$ regret are asymptotically optimal for all product revenue parameters.


Introduction
In this note, we consider the problem of assortment optimization under the multinomial logit (MNL) choice model with a capacity constraint on the size of the assortment and incomplete information about the model parameters. This problem has recently received considerable attention in the literature (see, e.g., [1][2][3][4][5][6][7]9,10]).
Two notable recent contributions are from Agrawal et al. [1,2], who construct decision policies based on Thompson Sampling and Upper Confidence Bounds, respectively, and show that the regret of these policies-the cumulative expected revenue loss compared with the benchmark of always offering an optimal assortment-is, up to logarithmic terms, bounded by a constant times √ , where denotes the number of products and denotes the length of the time horizon. These upper bounds are complemented by the recent work from Chen and Wang [3], who show that the regret of any policy is bounded from below by a positive constant times √ , implying that the policies by Agrawal et al. [1,2] are (up to logarithmic terms) asymptotically optimal.
The lower bound by Chen and Wang [3] is proven under the assumption that the product revenues are constant-that is, each product generates the same amount of revenue when sold. In practice, it often happens that different products have different marginal revenues, and it is a priori not completely clear whether the policies by Agrawal et al. [1,2] are still asymptotically optimal or that a lower regret can be achieved. In addition, Chen and Wang [3] assume that , the maximum number of products allowed in an assortment, is bounded by 1 4 · , but point out that this constant 1 4 can probably be increased. In this note, we settle this open question by proving a √ regret lower bound for any given vector of product revenues. This implies that policies with O( √ ) regret are asymptotically optimal regardless of the product revenue parameters. Furthermore, our result is valid for all < 1 2 , thereby confirming the intuition of Chen and Wang [3] that the constraint 1 4 is not tight.

Model and main result
We consider the problem of dynamic assortment optimization under the MNL choice model. In this model, the number of products is ∈ N. Henceforth, we abbreviate the set of products  is the set of possible histories. Then, conditionally on ℎ = ( 1 , 1 , . . . , −1 , −1 ), assortment has distribution ( · | ℎ), for all ℎ ∈ and all ∈ [ ]. Let E be the expectation operator under policy and preference vector ∈ V := [0, ∞) . The objective for the seller is to find a policy that maximizes the total accumulated revenue or, equivalently, minimizes the accumulated regret: where ( , ) is the expected revenue of assortment ⊆ [ ] under preference vector ∈ V: In addition, we define the worst-case regret: The main result, presented below, states that the regret of any policy can uniformly be bounded from below by a constant times √ .

Proof outline
The proof of Theorem 1 can be broken up into four steps. First, we define a baseline preference vector 0 ∈ V and we show that under 0 any assortment ∈ S is optimal. Second, for each ∈ A , we define a preference vector ∈ V by for some ∈ (0, 1]. For each such , we show that the instantaneous regret from offering a suboptimal assortment is bounded from below by a constant times the number of products | \ | not in ; cf. Lemma 1 below. This lower bound takes into account how much the assortments 1 , . . . , overlap with when the preference vector is . Third, let denote the number of times product ∈ [ ] is contained 1 , . . . , , that is, Then, we use the Kullback-Leibler (KL) divergence and Pinsker's inequality to upper bound the difference between the expected value of under and \{ } , see Lemma 2. Fourth, we apply a randomization argument over { : ∈ S }, we combine the previous steps, and we set accordingly to conclude the proof.
The novelty of this work is concentrated in the first two steps. The third and fourth step closely follow the work of Chen and Wang [3]. These last steps are included (1) because of slight deviations in our setup, (2) for the sake of completeness, and (3) since the proof techniques are extended to the case where / < 1/2. In the work of Chen and Wang [3], the lower bound is shown for / 1/4, but the authors already mention that this constraint can probably be relaxed. Our proof confirms that this is indeed the case.

Step 1: Construction of the baseline preference vector
Let := min ∈[ ] > 0 and define the constant Note that < . The baseline preference vector is formally defined as Now, the expected revenue for any ∈ A under 0 can be rewritten as The expression on the right-hand side is only maximized by assortments with maximal size | | = , in which case It follows that all assortments with size are optimal.

Step 2: Lower bound on the instantaneous regret of
For the second step, we bound the instantaneous regret under .
As a consequence, Proof. Fix ∈ S . First, note that since 1, for any ∈ A , it holds that Second, let * ∈ arg max ∈A ( , ) and * = ( * , ). By rewriting the inequality * ( , ) for all ∈ A , we find that for all ∈ A * ∈ ( − * ). Here, the first inequality is due to (3.3) and the second inequality follows from (3.4) with = . Next, note that since | | and | | = , we find that Now, term ( ) can be bounded from below as Here, at the final inequality, we used (3.5). Next, term ( ) can be bounded from above as . Now, for term ( ), we note that 0 for all ∈ [ ]. In addition, since ( * , 0 ) , * − ∈ * 1 + ∈ * − ∈ * 0 1 + ∈ * 0 1 1 This entails an upper bound for ( ). Term ( ) is bounded from above as Here, at the final inequality, we used (3.5) and the fact that 1. Now, we combine the upper bounds of ( ) and ( ) to find that It follows from (3.6) and (3.7) that Note that the constant 2 is positive if ( − ) 2 > 3 . This follows from = 2 /(3 + 2 ) since

Step 3: KL divergence and Pinsker's inequality
We denote the dependence of the expected value and the probability on the preference vector as E [ · ] and P (·) for ∈ A . In addition, we write \ instead of \{ }. The lemma below states an upper bound on the KL divergence of P and P \ and uses Pinsker's inequality to derive an upper bound on the absolute difference between the expected value of under and \ .
As a consequence, Proof. Let P and Q be arbitrary probability measures on ∪ {0}. It can be shown, see, for example, Lemma 3 from Chen and Wang [3], that where and are the probabilities of outcome under P and Q, respectively. We apply this result for and defined as for ∈ ∪ {0}. First, note that by (3.3), for all ∈ ∪ {0}, Now, we bound | − | from above for ∈ ∪ {0}. Note that for = 0 it holds that For ≠ , since 1, we find that Therefore, we conclude that KL(P ( · | ) || P \ ( · | )) ∈ ∪{0} Next, note that the entire probability measures P and P \ depend on . Then, as a consequence of the chain rule of the KL divergence, we find that KL(P || P \ ) where the step in (3.9) follows from, for example, Proposition 4.2 from Levin et al. [8] and we used Pinsker's inequality at the final inequality.

Step 4: Proving the main result
With all the established ingredients, we can finalize the proof of the lower bound on the regret.
Proof of Theorem 1. Since ∈ V for all ∈ S and by Lemma 1, we know that (3.10) We decompose ( ) into two terms: ( ) .
Recall that = / ∈ (0, 1/2). By summing over = \ instead of over , we bound ( ) from above by where the first inequality follows from ∈[ ] E [ ] , and the second inequality from Now, ( ) can be bounded by applying Lemma 2: